Samsung's HBM4 memory enters trial production, with mass production targeted for late 2025

Samsung Electronics has wrapped up the logic chip design for its next-generation High Bandwidth Memory (HBM4) and launched 4nm trial production, marking a big step toward mass production in the latter half of 2025. This updated schedule is about six months ahead of the original 2026 plan, potentially allowing Samsung to nab key orders for Nvidia’s future GPU platforms.

Word on the street is that Samsung’s storage division has already finished designing the logic chips, while its foundry division is hard at work with 4nm trial runs. The HBM4 production is happening at a specialized D1c facility, which uses an advanced 10nm DRAM process. Samsung has even skipped over the usual D1b development stage to speed things up.

Technical details shared at ISSCC 2024 show that HBM4 will offer a hefty performance boost, with data transfer rates of up to 2 TB/s—about 66 percent faster than HBM3E. It will also support a 2048-bit interface running at 6.4 GT/s and bump capacity up to 48 GB, which is 33 percent larger than the current generation.

Meanwhile, SK Hynix, Samsung’s biggest rival in the HBM space, is reportedly also fast-tracking its HBM4 development to hit the same 2025 target. Analysts at Hanwha Investment & Securities expect SK Hynix to keep its lead in market share and possibly be the first to ship HBM4 samples to customers.

Beyond Nvidia, Samsung is customizing HBM4 products for Microsoft and Meta. This memory tech plays a critical role in Nvidia’s upcoming “Rubin” GPU platform, slated for 2026, which will feature eight HBM4 chips and rely on TSMC’s 3nm process.

Samsung’s current lineup of HBM products is already riding a wave of momentum. Sales in Q3 2024 surged more than 70 percent month-over-month. HBM3E products, including 8-layer and 12-layer versions, have entered mass production and are expected to account for roughly half of Samsung’s total HBM sales by the end of 2024.